`fix(megatron): align on-policy resume offset and sampler seed with rollout`

[wzkk123]

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

Two independent dataloader bugs in on-policy (GRPO/GKD) Megatron training:

1. Resume offset. Megatron checkpoint state.consumed_train_samples counts
   optimized rollout rows (queries × n). The on-policy dataloader indexes
   unrepeated query rows. _prepare_dataloader fed the raw persisted offset into MegatronPretrainingRandomSampler, so after a resume with n-way generation the sampler skipped n times too far — training silently continued on the wrong part of the dataset.

2. Sampler seed. MegatronPretrainingRandomSampler seeded the epoch permutation with epoch alone, so every run (regardless of the configured data_seed) shared the same shuffle order, and data_seed had no effect on
   on-policy data ordering.

How to Reproduce / 如何复现

• _prepare_dataloader (base.py): for rlhf_type in {'grpo', 'gkd'} convert consumed_train_samples back to query rows (//= num_generations), failing closed with RuntimeError if the persisted value is not divisible by
  num_generations.
• MegatronPretrainingRandomSampler (batch_sampler.py): accept a seed parameter; use seed + epoch for the epoch permutation generator.
• _prepare_dataloader passes seed=args.data_seed (available from the base SFT arguments).

Additional Information / 补充信息

No response